DI-UMONS : Dépôt institutionnel de l’université de Mons

Recherche transversale
(titres de publication, de périodique et noms de colloque inclus)
2018-04-13 - Article/Dans un journal avec peer-review - Anglais - 14 page(s)

Kasmi Najlae, Mostapha Zbakh, Mahmoudi Sidi , Manneback Pierre , "WT_ DMDA New Scheduling strategy for Conjugate Gradient Solver on Heterogeneous" in International Journal of Autonomic Computing

  • Edition : Inderscience, Olney (United Kingdom)
  • Codes CREF : Informatique générale (DI1162), Imagerie médicale, radiologie, tomographie (DI3243)
  • Unités de recherche UMONS : Informatique, Logiciel et Intelligence artificielle (F114)
  • Instituts UMONS : Institut de Recherche en Technologies de l’Information et Sciences de l’Informatique (InforTech)
  • Centres UMONS : Centre de Recherche en Technologie de l’Information (CRTI)
Texte intégral :

Abstract(s) :

(Anglais) Heterogeneous systems which are composed of multiple CPUs and GPUs are more rnand more attractive as platforms for high performance computing. With thern evolution of General Purpose computation on GPU (GPGPU) and correspondingrn programming frameworks (OpenCL and CUDA), more applications are using GPUs rnas a co-processor to achieve performance that could not be accomplished usingrn just the traditional processors.However, the main problem is identifying which rntask or job should be allocated to a particular device.rnThe problem is even complicated due to the dissimilar computational power of rnthe CPU and the GPU. In this work we propose a new scheduling strategy WT_ dmdarnwhich aims to optimize the performance of the preconditioned conjugate gradientrnsolver, in CPU-GPU heterogeneous environment.We use StarPU runtime system to rnassess the efficiency of the approach on a computational platform consistingrnof three NVIDIA Fermi GPUs and twelve Intel CPUs. We show important speed up rn(up to 5.13) may be reached (relatively to default scheduler of StarPU)rnwhen processing large matrices and that the performance is advantageous rnwhen changing the granularity of tasks. An analysis and evaluation of rnthese results is discussed